Re: [PATCH v3] cpufreq: brcmstb-avs: fix imbalance of cpufreq policy refcount

From: Viresh Kumar <viresh.kumar@linaro.org>
To: chenqiwu <qiwuchen55@gmail.com>
Cc: mmayer@broadcom.com, rjw@rjwysocki.net, f.fainelli@gmail.com,
	bcm-kernel-feedback-list@broadcom.com, linux-pm@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org,
	linux-kernel@vger.kernel.org, chenqiwu <chenqiwu@xiaomi.com>
Subject: Re: [PATCH v3] cpufreq: brcmstb-avs: fix imbalance of cpufreq policy refcount
Date: Mon, 20 Jan 2020 12:00:04 +0530	[thread overview]
Message-ID: <20200120063004.zzhep35vfl3urndd@vireshk-i7> (raw)
In-Reply-To: <20200120062756.GA5802@cqw-OptiPlex-7050>

On 20-01-20, 14:27, chenqiwu wrote:
> On Mon, Jan 20, 2020 at 11:51:26AM +0530, Viresh Kumar wrote:
> > On 20-01-20, 14:13, chenqiwu wrote:
> > > On Mon, Jan 20, 2020 at 11:31:34AM +0530, Viresh Kumar wrote:
> > > > On 20-01-20, 13:58, chenqiwu wrote:
> > > > > On Mon, Jan 20, 2020 at 11:02:50AM +0530, Viresh Kumar wrote:
> > > > > > On 19-01-20, 15:09, qiwuchen55@gmail.com wrote:
> > > > > > > From: chenqiwu <chenqiwu@xiaomi.com>
> > > > > > > 
> > > > > > > brcm_avs_cpufreq_get() calls cpufreq_cpu_get() to get the cpufreq policy,
> > > > > > > meanwhile, it also increments the kobject reference count to mark it busy.
> > > > > > > However, a corresponding call of cpufreq_cpu_put() is ignored to decrement
> > > > > > > the kobject reference count back, which may lead to a potential stuck risk
> > > > > > > that the cpuhp thread deadly waits for dropping of kobject refcount when
> > > > > > > cpufreq policy free.
> > > > > > > 
> > > > > > > For fixing this bug, cpufreq_get_policy() is referenced to do a proper
> > > > > > > cpufreq_cpu_get()/cpufreq_cpu_put() and fill a policy copy for the user.
> > > > > > > If the policy return NULL, we just return 0 to hit the code path of
> > > > > > > cpufreq_driver->get.
> > > > > > > 
> > > > > > > Signed-off-by: chenqiwu <chenqiwu@xiaomi.com>
> > > > > > > ---
> > > > > > >  drivers/cpufreq/brcmstb-avs-cpufreq.c | 12 ++++++++++--
> > > > > > >  1 file changed, 10 insertions(+), 2 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/drivers/cpufreq/brcmstb-avs-cpufreq.c b/drivers/cpufreq/brcmstb-avs-cpufreq.c
> > > > > > > index 77b0e5d..ee0d404 100644
> > > > > > > --- a/drivers/cpufreq/brcmstb-avs-cpufreq.c
> > > > > > > +++ b/drivers/cpufreq/brcmstb-avs-cpufreq.c
> > > > > > > @@ -452,8 +452,16 @@ static bool brcm_avs_is_firmware_loaded(struct private_data *priv)
> > > > > > >  
> > > > > > >  static unsigned int brcm_avs_cpufreq_get(unsigned int cpu)
> > > > > > >  {
> > > > > > > -	struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
> > > > > > 
> > > > > > Why can't we just add a corresponding cpufreq_cpu_put() instead of all this ?
> > > > > > 
> > > > > 
> > > > > cpufreq_get_policy() does a proper cpufreq_cpu_get()/cpufreq_cpu_put(),
> > > > > meanwhile fills a policy copy for the user. It equals to using
> > > > > cpufreq_cpu_get() and a corresponding cpufreq_cpu_put() around access
> > > > > to the policy pointer. I think both methods are fine here.
> > > > > What do you think?
> > > > 
> > > > cpufreq_get_policy() does an extra memcpy as well, which isn't required at all
> > > > in your case.
> > > > 
> > > > -- 
> > > > viresh
> > > 
> > > Huha..Do you worry about the race conditon with cpufreq policy free path?
> > 
> > No. I just worry about an unnecessary memcpy, nothing else.
> >
> Is there any question about this extra memcpy?

What do you mean by that?

The whole point I am trying to make is that for your specific case, doing an
explicit cpufreq_cpu_get() and cpufreq_cpu_put() is far more efficient than
calling cpufreq_get_policy() which has a different purpose and usecase.

-- 
viresh